Knowledge cutoff isn't particularly relevant for search. I want an LLM that has access to a good search tool and knows how to use it.
Gemini should be excellent here - it should have access to the best search index out there.
But... it doesn't show me what it's searching for. This is an absolute show-stopper for me: I need to know what the LLM is searching for in order to evaluate if it is likely to find the right information for me.
ChatGPT gets this right: it shows me the search terms it's using, then shows me the articles from the results that it used to generate the response.
Until a few weeks ago I still didn't use it much, because inevitably when it ran a search I would shout at my computer "No, don't search for that! You'll get crap results".
This changed with the recent release of o3 and o4-mini. For the first time I feel like I have access to a model with genuinely good taste in searching - it picks good initial searches, then revises those searches based on the incoming results.
Gemini should be excellent here - it should have access to the best search index out there.
But... it doesn't show me what it's searching for. This is an absolute show-stopper for me: I need to know what the LLM is searching for in order to evaluate if it is likely to find the right information for me.
ChatGPT gets this right: it shows me the search terms it's using, then shows me the articles from the results that it used to generate the response.
Until a few weeks ago I still didn't use it much, because inevitably when it ran a search I would shout at my computer "No, don't search for that! You'll get crap results".
This changed with the recent release of o3 and o4-mini. For the first time I feel like I have access to a model with genuinely good taste in searching - it picks good initial searches, then revises those searches based on the incoming results.
I wrote about that recently: https://simonwillison.net/2025/Apr/21/ai-assisted-search/